What are we measuring when we measure risk attitudes?
A difficult task with crucial relevance
and by the way, what is risk?
The act of implementing a goal-directed option qualifies as an instance of risk taking whenever two things are true: (a) the behavior in question could lead to more than one outcome and (b) some of these outcomes are undesirable or even dangerous. In essence, then, risk taking involves the implementation of options that could lead to negative consequences.
(Byrnes et al 1999)
risk loosely defined as probability of harm
focus on questionnaires and intuitive tasks
Metrics of success: convergent validity + predictive validity
decisions given a probability distribution over outcomes
if probability and outcomes known: risk
if only oucomes known: ambiguity
if both unknown: knightian uncertainty
risk formally defined as uncertainty over outcomes
focus on decontextualized tasks (and questionnaires)
Metric of success: internal validity (task \(\iff\) theory)
How likely are you to take risks in general, one a scale from 0 (not taking any risks) to 10 (taking many risks)?
Domain Specific Risk Taking Scale
Examples:
A meta-analysis of Risk elicitation tasks
elicited risk atitudes: tasks and questionnaires
convergent validity: correlation among tasks
convergent validity: correlation among questionnaires
predictive validity: correlation task \(\iff\) questionnaires
\(u(x) = x^r\)
low consistency across tasks
surprisingly, low consistency also within tasks
but heterogeneity by task is large
only result that holds: most people are risk averse
possible explanation: between-subjects variation.
better consistency across samples
a tendency to report ‘in the middle’
we do not really know what those numbers mean
we replicate Slovic 1962 (!!)
no correlation higher than .35
when transalitng into r things get worse
low correlations with questionnaires
across questionnaires and tasks
Beauchamp et al JRU 2016: questionnaires are rather predictive
task-specific bias
noise
risk perception
theory
task-specific bias
(noise)
risk perception
(theory)
noisy preference + one-shot choices \(\Rightarrow\) noisy data
cognitive limits \(\Rightarrow\) limited understanding
task-specific bias?
(this work: Crosetto and Filippin, ExEc 2015)
How does the mere mechanics of each task affect the outcome?
Simulation exercise:
A good task should be able to recreate the starting distribution, if no error.
3 types of simulations:
deterministic
random parameter model \(\Rightarrow\) models fuzzy preferences
random agents \(\Rightarrow\) models frame effects
is there a task-specific bias? yes
does it account for all differences? no
is this the only way to take noise into account? no
economists assume subjects share the same risk definition
namely:
but subjects think of risk as probability of a loss
if:
then:
send your data – paolo.crosetto@inrae.fr
github: (https://github.com/paolocrosetto/METARET)
shiny app: (https://paolocrosetto.shinyapps.io/METARET/)